All Updates
Page 740 of 1667
May 5, 2026
US-China Aces: Tech Strengths Clash
Contrasts US VC-driven R&D innovation with China's state-backed tech diffusion strategy. AI accelerates diffusion by cutting engineer needs for improvements. Geopolitics echoes historical rail/AI sovereignty battles.
Random Walks Outperform Precise Design in Evolution
Artem Kaznatcheev models biological evolution as algorithms navigating fitness landscapes. Additive landscapes enable easy gradient ascent (O(1) complexity), while epistatic ones form maze-like paths with exponential difficulty (O(e^n)). This explains why some traits evolve readily, others remain intractable even over planetary timescales.
OpenAI's 2027 AI Agent Phone Rumored
OpenAI may launch its first AI-first smartphone in 2027. It focuses on on-device intelligence with advanced processors and next-gen memory. This enables smarter, more autonomous user interactions.
DeepMind UK Workers Unionize Over Military Deal
Google DeepMind workers in the UK have voted to unionize due to concerns over a recent deal with the US military. They sent a letter requesting recognition of the Communication Workers Union and Unite the Union. Issues cited include the Iran war and the Pentagon's feud with Anthropic.
Qualcomm Vet Leads Intel Physical AI
Intel appoints ex-Qualcomm exec Alex Katouzian as EVP and GM of Client Computing and Physical AI Group starting May. Pushkar Ranade is named permanent CTO, focusing on quantum, neuromorphic computing, photonics, and new materials. Both report to CEO Pat Gelsinger.
China Mobile Launches AI-eSIM
China Mobile will launch AI-eSIM at the 2026 Mobile Cloud Conference in Suzhou from May 7-9. The product enables real-time cloud model scheduling for autonomous device thinking and instant responses. Targeted at AI toys, smart wearables, and other terminals.
China Approves First L3 Autonomous Cars
China's MIIT approved the first two L3 conditional autonomous driving vehicles for production and road use by December 2025. L4 pilots are advancing in cities like Guangzhou and Hainan for sanitation and tourism. Autonomous driving is positioned as key to breaking auto industry 'involution' via new services and ecosystems.
Architect AI Agents, Not Manage People
Expert Sun Tianzhu urges entrepreneurs to shift from human management to building AI agent-centric organizations amid rapid agent iteration. AI disrupts short drama and software sectors, enabling one-person teams. Future firms need AI architects defining vision atop agent execution layers.
UK Kids Bypass Age Checks with Fake Beards
Internet Matters report reveals 32% of UK children aged 9-16 admitted bypassing age verification on social media and adult sites. Methods include fake beards, VPNs, and parent accounts. The Online Safety Act shows limited effectiveness per survey of 1270 kids and parents.
Huang: AI Creates Jobs, Not Unemployment
Nvidia CEO Jensen Huang refutes AI-driven unemployment fears, calling AI a massive job creator. He expressed this during a Milken Institute event dialogue with MSNBC's Becky Quick.
OpenAI Eyes Robotics-Hardware Split Pre-IPO
OpenAI CEO Sam Altman considered spinning off robotics and consumer hardware units to avoid burdening core AI business and enable independent funding. The plan was shelved over balance sheet concerns but may revive as Alphabet-style holding. Units operate secretly, with recent $6B io acquisition and Apple talent poaching.
Apple Eyes Intel, Samsung Chip Backup
Apple is in early talks with Intel and Samsung to produce device main processors as TSMC alternatives. Supply disruptions from AI data center boom and Mac demand shortages prompt diversification. No orders placed yet; concerns remain over non-TSMC tech.
xAI's 550K GPUs Run at Just 11% Utilization
xAI's Memphis Colossus cluster of 550k Nvidia H100/H200 GPUs achieves only 11% MFU due to scaling bottlenecks in networking and data pipelines. Industry giants like Meta (43%) and Google (46%) fare better via optimized stacks. xAI targets 50% via software fixes and eyes agentic AI leasing.
AI Threatens Software Debt Recovery
AI disruptions are threatening private credit firms' potential recovery rates in the software sector. Davidson Kempner Capital Management LP CIO Tony Yoseloff highlighted this risk.
TUR-DPO: Topology-Aware DPO Upgrade
TUR-DPO is a new variant of Direct Preference Optimization that incorporates lightweight reasoning topologies and uncertainty signals for better LLM alignment. It improves judge win-rates, faithfulness, and calibration across math reasoning, QA, summarization, and dialogue benchmarks on 7-8B models. The method remains RL-free, matching PPO on reasoning tasks while preserving simplicity.
Tool-Use Tax Exposed in LLM Agents
Tool-augmented reasoning in LLM agents does not always outperform native Chain-of-Thought (CoT) under semantic distractors. A Factorized Intervention Framework reveals the 'tool-use tax' from prompt formatting and tool-calling protocols often outweighs tool gains. G-STEP, a lightweight inference-time gate, partially mitigates these protocol-induced errors.
TokenArena: AI Inference Benchmark Launch
TokenArena introduces a continuous benchmark evaluating AI inference at endpoint level across speed, price, energy, context, and quality. It analyzes 78 endpoints from 12 model families, revealing up to 12.5-point accuracy differences and 6.2x variations in energy per correct answer. The framework, schema, and leaderboard are released openly under CC BY 4.0.
TADI: Agentic LLM for Drilling Intelligence
TADI is an agentic AI system integrating 1,759 daily drilling reports, WITSML data, and production records from Equinor Volve Field into DuckDB and ChromaDB stores. It uses 12 domain-specific tools orchestrated by an LLM for multi-step evidence gathering across structured and narrative data. The reproducible implementation parses all DDR XML files error-free and introduces the Evidence Grounding Score for evaluation.
LOCA Explains Specific LLM Jailbreaks
Researchers introduce LOCA, a method providing local, causal explanations for jailbreak success in safety-trained LLMs by identifying minimal interpretable changes in intermediate representations to induce refusal. Evaluated on Gemma and Llama models using a large jailbreak benchmark, LOCA achieves refusal with an average of six changes, outperforming prior methods that often fail even after 20 changes. This advances mechanistic understanding of jailbreaks beyond global explanations.
Hamiltonian World Models for Reliable Embodied AI
World models are key for embodied intelligence but current approaches lack physical reliability and action control. This paper proposes Hamiltonian World Models, encoding observations into latent phase space and evolving states via Hamiltonian dynamics with control, dissipation, and residuals. It enables interpretable, stable predictions for robotics and RL planning.